Model deprecation
A provider announcing that a model will be retired, followed by a retirement date after which requests to it fail.
Why it matters. Every hosted model you depend on will eventually be retired, so plan migrations and keep tests ready.
How it works
Deprecation is the process of retiring a model: the provider announces it, usually names a replacement, and sets a retirement date after which calls to that model ID return errors. Between announcement and retirement the model normally keeps working.
Notice periods and policies differ by provider, and cloud platforms may set their own dates. Many providers publish a deprecations page listing dates. Open-weights models you host yourself are not retired by anyone, though hosts may stop serving them.
Track the deprecation pages of every provider you use and keep an evaluation set ready, so a migration is a test run plus a config change. The replacement may differ in price, tokenizer, defaults and supported parameters.
Worked example
Illustrative: a provider announces on 1 March that a model retires on 1 September. A team with a ready evaluation suite spends a week comparing the replacement; a team without one finds out on 1 September when production requests start failing.
Common mistakes
- Not following provider deprecation notices or checking the deprecations page.
- Assuming the replacement accepts the same parameters and produces the same output format.
- Forgetting that cloud platforms may retire a model on different dates from the maker.
Across the catalog today
Next retirement dates in the catalog
- Claude Haiku 4.5Active (legacy)Not before 15 Oct 2026
- Claude Opus 4.5Active (legacy)Not before 24 Nov 2026
- Claude Sonnet 4.5Deprecated30 Nov 2026
- Claude Opus 4.6Active (legacy)Not before 5 Feb 2027
- Claude Sonnet 4.6Active (legacy)Not before 17 Feb 2027
- Claude Opus 4.7Active (legacy)Not before 16 Apr 2027
“Not before” dates are the maker's minimum guarantee, not a scheduled shutdown.
See it on a real card: Overview on Claude Opus 5.5.